Random bin splits gave higher validation correlation for a fixed decoder of heard speech amplitude in ten SEEG recordings

We tested a fixed decoder that used eight SEEG pairs to read heard speech amplitude. The public source was OpenNeuro ds004703 v1.1.0. We compared random bin splits with complete-passage splits in ten people. Training row counts were equal. Mean passage correlation was 0.161 with random bins and -0.015 with complete passages. The paired difference was 0.176. Its descriptive 95 percent person bootstrap interval was 0.147 to 0.208. All ten differences were positive. Shifted audio gave a similar difference of 0.175. All random-bin validation rows shared gamma signal samples with fitting rows. Complete-passage validation rows shared none. The later-passage score was -0.012. These higher random-bin scores did not establish transfer to new passages. The test does not isolate sample overlap from shared passage context. The package preserves input features, source hashes, splits, targets, models, predictions, three figures, and reproduction code.

stereoelectroencephalography · speech amplitude · brain-computer interface · validation · random bin split · passage split · ridge regression · sample overlap